How do I cut AI costs when using Claude or Codex all day?

I jumped to the $200 a month Codex plan and used 80% of it in 36 hours. I need better AI cost discipline, and I say so openly. Here's what I've learned from my own bills and from builders I talk to.

Are the $200 a month plans worth it?

Usually, because they're subsidized. A co-founder and I compared notes: I cited roughly $7,000 to $8,000 of Claude usage and about $14,000 of ChatGPT usage, in API-equivalent terms, under flat monthly plans. Those are numbers I'd heard, not an audit, but flat plans hide a lot of usage.

I pay $200 for each of the two. At one point I was running five products at the same time, at least 16 hours a day. Without the flat plans I'd guess I'd be spending five to ten grand a month. I expect them to lower this at some point, which is why now is the time to build. That only matters if you actually have things to build.

Everybody should start on the $20 plan and see how it goes. You'll get cut off quickly on the coding tools. I ran out of Codex on $20, upgraded to $100, and was close to the weekly limit within 48 hours, partly because I left reasoning on high or extra high for jobs that didn't need it.

Should agents use API keys or your subscription?

Subscription first, every time. My agents default to my flat-rate ChatGPT and Claude accounts. A pay-per-use key for ChatGPT or Claude is banned in my systems without my explicit approval.

"Don't ever, ever assume we're doing a pay-per-use key for ChatGPT or Claude. That is absolutely banned without 100% approval."

When an API truly is needed, it goes through OpenRouter to a cheap open-source model like GLM or DeepSeek. For that kind of background work, Sonnet is way too expensive. The model still has to handle vision and files, because some can't. And put a cap on it: I found my OpenRouter limit was set to forever, not monthly, and changed it to $100 a month.

For a team writing content, the API can be the cheaper route. If your prompts live in your own software, API costs might run 10 to 20% of what you pay for ChatGPT seats. When my ChatGPT bill was $150 a month, I was about ready to dump it for exactly that reason. Each content type gets its own saved prompt, so anyone can paste a business link and get an ad in the house style.

Which model should I use for what?

Job	Model
Planning a build, setting up systems	Top model
Implementing the plan	Cheaper model
Everyday chats on a $20 plan	Sonnet or Haiku
Newsletter research prompts	Opus
Background API work	Open-source model via OpenRouter

The bigger model uses more of your allowance per task. My made-up example: if you could do 100 chats with Opus, you could do 400 with Sonnet. That's why Sonnet is the default across my business for simple work. My media buyer pushed back that Sonnet doesn't research as thoroughly, and I agreed, so research gets Opus. On the $200 plan I personally just leave it on Opus.

A developer friend told me he codes with Haiku, the cheapest Claude model, and it works well. That's crazy to me, but it is. You just have to know when you need each model.

How should a system move from cheap models to expensive ones?

Start cheap and escalate. If a cheaper model can't get something done, or does it wrong, the system upgrades to a stronger model before anyone bothers me. My order is free or cheap models first, then GLM, then Claude or the better Codex models. Only if those fail does it come to me.

Free models are worth using while they last. One cheap coding plan I use gives about $60 a month of tokens plus free models that rotate, and one of them was good at bug fixing.

What burns the most tokens?

• Images and video. Those are the main token burners.
• Effort set too high. High or extra-high reasoning on a simple bug fix is how I got close to a weekly limit in two days.
• Agents looping on tiny details. I watched one agent go through 16 rounds of revisions with another AI reviewer over tiny, unimportant things. I suspect some of the standing instructions I added to help were making it worse. The fix is tuning the tools to understand what I want with the least overhead.

What about the cloud bills behind the AI?

They sneak up on you. GitHub, Vercel and Supabase all started charging me a lot. My hard rule now: any change an agent makes that could raise my cost has to say how much, with a reasonable monthly estimate.

The classic trap is a worker that runs every 5 minutes instead of every 15. That's three times the runs on services that bill by usage. If a change only adds $5 or $10, I don't care. Tripling the bill is a huge problem.

Here's roughly what my stack looked like in October 2026: $200 for ChatGPT Pro, $200 for Claude Max, $10 for a cheap coding-model plan, $30 for an agent tool, $80 for Supabase, about $100 for Vercel and $25 for Slack, plus OpenRouter. It's a valuable setup, but it only pays for itself if the agents bring in leads. So I ask my system for a full audit: every model used, cost and tokens, a plain explanation of what each piece is worth, and smart changes to get the best value.

What else cuts the bill?

Two builders I talk to run lean. These are their numbers, not mine:

• One peer runs GitHub Copilot at about $39 a month on an older Claude model at medium reasoning and says he never hit a token limit in two-day sessions. He also says structured, dashed-list prompts use fewer tokens.
• Another contact codes 10 to 12 hours a day on a $40 Copilot plan, using frontier models for planning only.

And the moves that work for me:

1. Move repeat work off expensive reasoning. When my AI-plan capacity ran dry in August 2026, the advice was to move deterministic execution, the same steps every time, away from expensive reasoning models.
2. Stay under free tiers for tools. When I gave a research agent Exa and Firecrawl, I made sure it stayed under their free tiers: research tools for AI agents.
3. Bundle the small subscriptions. Lenny's bundle gets you a pile of AI tools for roughly $20 or $30 a month. A year of the dictation tool alone is worth about $240.
4. Split planning from building. That's the biggest lever: plan with one AI, build with another.